Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning
The paper turns physical scenes into executable code, then uses those verified simulations to train models on quantitative reasoning.
Code-as-World represents composition, dynamics, and appearance as runnable world hypotheses. An agent builds these from language or video, executes and renders them, then checks and refines the result. The authors use the verified worlds as scalable supervision for vision-language models. They report state-of-the-art results on QuantiPhy and say Code-as-World-VL beats leading proprietary models. HF Daily Papers' note
Code-as-World represents composition, dynamics, and appearance as runnable world hypotheses. An agent builds these from language or video, executes and renders them, then checks and refines the result. The authors use the verified worlds as scalable supervision for vision-language models. They report state-of-the-art results on QuantiPhy and say Code-as-World-VL beats leading proprietary models. HF Daily Papers' note
score 5